🏆 Winner Solution for NTIRE 2025 RAW Image Super-Resolution Challenge

Tianyu Zhang^1♠️, Xin Luo^1♠️, Yeda Chen², Dong Liu¹

¹ University of Science and Technology of China, Hefei, China

² Shanghai Shuangshen Information Technology Co., Ltd.

^♠️ Equal contribution

📌 Introduction

This repository contains the official implementation of our award-winning solution for the NTIRE 2025 RAW Image Super-Resolution Challenge. Our method achieves state-of-the-art performance with enhanced degradation modeling and efficient architecture design.

📦 Installation

# create a virtual environment [Recommended but optional]
conda create -n lite_rawformer python=3.11
source activate lite_rawformer

# Install all necessary dependencies
# In root/
bash install.sh

# 如果你是来自中国大陆的用户，可以使用下面的命令从国内源进行安装。
# 下面的命令默认使用CUDA 12.4，如果需要其他版本，请自行修改文件第一行。
bash install_zh.sh

🚀 Inference

• Download our pretrained models, and place the LiteRAWFormer folder in pretrained_models/
• Download the validation inputs, and place it in datasets/RAWSR/
• Running inference with following command, the results are saved in results/RAWSR/val_out

  # in root directory
  bash inference.sh

• To evaluate the results, you should register NTIRE 2025 RAW Restoration Challenge and upload your results to the platform.

⛵ Training

• Download the Training set, and place it in datasets/RAWSR/
• Running training with following command.

  # in root directory
  bash train.sh # use all available GPUs
  CUDA_VISIBLE_DEVICES=0,1 bash train.sh # use the first two GPUs

📧 Contact

If you have any questions, please open an issue (the recommended way) or contact us via

• [email protected]
• [email protected]

License

This work is licensed under MIT license. See the LICENSE for details.

Acknowledgement

Our repository builds upon the excellent framework provided by accelerate, and our architecture are inspired by RBSFormer.